1 //===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the SampleProfileLoader transformation. This pass 11 // reads a profile file generated by a sampling profiler (e.g. Linux Perf - 12 // http://perf.wiki.kernel.org/) and generates IR metadata to reflect the 13 // profile information in the given profile. 14 // 15 // This pass generates branch weight annotations on the IR: 16 // 17 // - prof: Represents branch weights. This annotation is added to branches 18 // to indicate the weights of each edge coming out of the branch. 19 // The weight of each edge is the weight of the target block for 20 // that edge. The weight of a block B is computed as the maximum 21 // number of samples found in B. 22 // 23 //===----------------------------------------------------------------------===// 24 25 #include "llvm/ADT/DenseMap.h" 26 #include "llvm/ADT/SmallPtrSet.h" 27 #include "llvm/ADT/SmallSet.h" 28 #include "llvm/ADT/StringRef.h" 29 #include "llvm/Analysis/LoopInfo.h" 30 #include "llvm/Analysis/PostDominators.h" 31 #include "llvm/IR/Constants.h" 32 #include "llvm/IR/DebugInfo.h" 33 #include "llvm/IR/DiagnosticInfo.h" 34 #include "llvm/IR/Dominators.h" 35 #include "llvm/IR/Function.h" 36 #include "llvm/IR/InstIterator.h" 37 #include "llvm/IR/Instructions.h" 38 #include "llvm/IR/LLVMContext.h" 39 #include "llvm/IR/MDBuilder.h" 40 #include "llvm/IR/Metadata.h" 41 #include "llvm/IR/Module.h" 42 #include "llvm/Pass.h" 43 #include "llvm/ProfileData/SampleProfReader.h" 44 #include "llvm/Support/CommandLine.h" 45 #include "llvm/Support/Debug.h" 46 #include "llvm/Support/ErrorOr.h" 47 #include "llvm/Support/Format.h" 48 #include "llvm/Support/raw_ostream.h" 49 #include "llvm/Transforms/IPO.h" 50 #include "llvm/Transforms/Utils/Cloning.h" 51 #include <cctype> 52 53 using namespace llvm; 54 using namespace sampleprof; 55 56 #define DEBUG_TYPE "sample-profile" 57 58 // Command line option to specify the file to read samples from. This is 59 // mainly used for debugging. 60 static cl::opt<std::string> SampleProfileFile( 61 "sample-profile-file", cl::init(""), cl::value_desc("filename"), 62 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden); 63 static cl::opt<unsigned> SampleProfileMaxPropagateIterations( 64 "sample-profile-max-propagate-iterations", cl::init(100), 65 cl::desc("Maximum number of iterations to go through when propagating " 66 "sample block/edge weights through the CFG.")); 67 static cl::opt<unsigned> SampleProfileRecordCoverage( 68 "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"), 69 cl::desc("Emit a warning if less than N% of records in the input profile " 70 "are matched to the IR.")); 71 static cl::opt<unsigned> SampleProfileSampleCoverage( 72 "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"), 73 cl::desc("Emit a warning if less than N% of samples in the input profile " 74 "are matched to the IR.")); 75 static cl::opt<double> SampleProfileHotThreshold( 76 "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"), 77 cl::desc("Inlined functions that account for more than N% of all samples " 78 "collected in the parent function, will be inlined again.")); 79 static cl::opt<double> SampleProfileGlobalHotThreshold( 80 "sample-profile-global-hot-threshold", cl::init(30), cl::value_desc("N"), 81 cl::desc("Top-level functions that account for more than N% of all samples " 82 "collected in the profile, will be marked as hot for the inliner " 83 "to consider.")); 84 static cl::opt<double> SampleProfileGlobalColdThreshold( 85 "sample-profile-global-cold-threshold", cl::init(0.5), cl::value_desc("N"), 86 cl::desc("Top-level functions that account for less than N% of all samples " 87 "collected in the profile, will be marked as cold for the inliner " 88 "to consider.")); 89 90 namespace { 91 typedef DenseMap<const BasicBlock *, uint64_t> BlockWeightMap; 92 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap; 93 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge; 94 typedef DenseMap<Edge, uint64_t> EdgeWeightMap; 95 typedef DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>> 96 BlockEdgeMap; 97 98 /// \brief Sample profile pass. 99 /// 100 /// This pass reads profile data from the file specified by 101 /// -sample-profile-file and annotates every affected function with the 102 /// profile information found in that file. 103 class SampleProfileLoader : public ModulePass { 104 public: 105 // Class identification, replacement for typeinfo 106 static char ID; 107 108 SampleProfileLoader(StringRef Name = SampleProfileFile) 109 : ModulePass(ID), DT(nullptr), PDT(nullptr), LI(nullptr), Reader(), 110 Samples(nullptr), Filename(Name), ProfileIsValid(false), 111 TotalCollectedSamples(0) { 112 initializeSampleProfileLoaderPass(*PassRegistry::getPassRegistry()); 113 } 114 115 bool doInitialization(Module &M) override; 116 117 void dump() { Reader->dump(); } 118 119 const char *getPassName() const override { return "Sample profile pass"; } 120 121 bool runOnModule(Module &M) override; 122 123 void getAnalysisUsage(AnalysisUsage &AU) const override { 124 AU.setPreservesCFG(); 125 } 126 127 protected: 128 bool runOnFunction(Function &F); 129 unsigned getFunctionLoc(Function &F); 130 bool emitAnnotations(Function &F); 131 ErrorOr<uint64_t> getInstWeight(const Instruction &I) const; 132 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB) const; 133 const FunctionSamples *findCalleeFunctionSamples(const CallInst &I) const; 134 const FunctionSamples *findFunctionSamples(const Instruction &I) const; 135 bool inlineHotFunctions(Function &F); 136 bool emitInlineHints(Function &F); 137 void printEdgeWeight(raw_ostream &OS, Edge E); 138 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const; 139 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB); 140 bool computeBlockWeights(Function &F); 141 void findEquivalenceClasses(Function &F); 142 void findEquivalencesFor(BasicBlock *BB1, 143 SmallVector<BasicBlock *, 8> Descendants, 144 DominatorTreeBase<BasicBlock> *DomTree); 145 void propagateWeights(Function &F); 146 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge); 147 void buildEdges(Function &F); 148 bool propagateThroughEdges(Function &F); 149 void computeDominanceAndLoopInfo(Function &F); 150 unsigned getOffset(unsigned L, unsigned H) const; 151 void clearFunctionData(); 152 153 /// \brief Map basic blocks to their computed weights. 154 /// 155 /// The weight of a basic block is defined to be the maximum 156 /// of all the instruction weights in that block. 157 BlockWeightMap BlockWeights; 158 159 /// \brief Map edges to their computed weights. 160 /// 161 /// Edge weights are computed by propagating basic block weights in 162 /// SampleProfile::propagateWeights. 163 EdgeWeightMap EdgeWeights; 164 165 /// \brief Set of visited blocks during propagation. 166 SmallPtrSet<const BasicBlock *, 128> VisitedBlocks; 167 168 /// \brief Set of visited edges during propagation. 169 SmallSet<Edge, 128> VisitedEdges; 170 171 /// \brief Equivalence classes for block weights. 172 /// 173 /// Two blocks BB1 and BB2 are in the same equivalence class if they 174 /// dominate and post-dominate each other, and they are in the same loop 175 /// nest. When this happens, the two blocks are guaranteed to execute 176 /// the same number of times. 177 EquivalenceClassMap EquivalenceClass; 178 179 /// \brief Dominance, post-dominance and loop information. 180 std::unique_ptr<DominatorTree> DT; 181 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT; 182 std::unique_ptr<LoopInfo> LI; 183 184 /// \brief Predecessors for each basic block in the CFG. 185 BlockEdgeMap Predecessors; 186 187 /// \brief Successors for each basic block in the CFG. 188 BlockEdgeMap Successors; 189 190 /// \brief Profile reader object. 191 std::unique_ptr<SampleProfileReader> Reader; 192 193 /// \brief Samples collected for the body of this function. 194 FunctionSamples *Samples; 195 196 /// \brief Name of the profile file to load. 197 StringRef Filename; 198 199 /// \brief Flag indicating whether the profile input loaded successfully. 200 bool ProfileIsValid; 201 202 /// \brief Total number of samples collected in this profile. 203 /// 204 /// This is the sum of all the samples collected in all the functions executed 205 /// at runtime. 206 uint64_t TotalCollectedSamples; 207 }; 208 209 class SampleCoverageTracker { 210 public: 211 SampleCoverageTracker() : SampleCoverage(), TotalUsedSamples(0) {} 212 213 bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset, 214 uint32_t Discriminator, uint64_t Samples); 215 unsigned computeCoverage(unsigned Used, unsigned Total) const; 216 unsigned countUsedRecords(const FunctionSamples *FS) const; 217 unsigned countBodyRecords(const FunctionSamples *FS) const; 218 uint64_t getTotalUsedSamples() const { return TotalUsedSamples; } 219 uint64_t countBodySamples(const FunctionSamples *FS) const; 220 void clear() { 221 SampleCoverage.clear(); 222 TotalUsedSamples = 0; 223 } 224 225 private: 226 typedef std::map<LineLocation, unsigned> BodySampleCoverageMap; 227 typedef DenseMap<const FunctionSamples *, BodySampleCoverageMap> 228 FunctionSamplesCoverageMap; 229 230 /// Coverage map for sampling records. 231 /// 232 /// This map keeps a record of sampling records that have been matched to 233 /// an IR instruction. This is used to detect some form of staleness in 234 /// profiles (see flag -sample-profile-check-coverage). 235 /// 236 /// Each entry in the map corresponds to a FunctionSamples instance. This is 237 /// another map that counts how many times the sample record at the 238 /// given location has been used. 239 FunctionSamplesCoverageMap SampleCoverage; 240 241 /// Number of samples used from the profile. 242 /// 243 /// When a sampling record is used for the first time, the samples from 244 /// that record are added to this accumulator. Coverage is later computed 245 /// based on the total number of samples available in this function and 246 /// its callsites. 247 /// 248 /// Note that this accumulator tracks samples used from a single function 249 /// and all the inlined callsites. Strictly, we should have a map of counters 250 /// keyed by FunctionSamples pointers, but these stats are cleared after 251 /// every function, so we just need to keep a single counter. 252 uint64_t TotalUsedSamples; 253 }; 254 255 SampleCoverageTracker CoverageTracker; 256 257 /// Return true if the given callsite is hot wrt to its caller. 258 /// 259 /// Functions that were inlined in the original binary will be represented 260 /// in the inline stack in the sample profile. If the profile shows that 261 /// the original inline decision was "good" (i.e., the callsite is executed 262 /// frequently), then we will recreate the inline decision and apply the 263 /// profile from the inlined callsite. 264 /// 265 /// To decide whether an inlined callsite is hot, we compute the fraction 266 /// of samples used by the callsite with respect to the total number of samples 267 /// collected in the caller. 268 /// 269 /// If that fraction is larger than the default given by 270 /// SampleProfileHotThreshold, the callsite will be inlined again. 271 bool callsiteIsHot(const FunctionSamples *CallerFS, 272 const FunctionSamples *CallsiteFS) { 273 if (!CallsiteFS) 274 return false; // The callsite was not inlined in the original binary. 275 276 uint64_t ParentTotalSamples = CallerFS->getTotalSamples(); 277 if (ParentTotalSamples == 0) 278 return false; // Avoid division by zero. 279 280 uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples(); 281 if (CallsiteTotalSamples == 0) 282 return false; // Callsite is trivially cold. 283 284 double PercentSamples = 285 (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0; 286 return PercentSamples >= SampleProfileHotThreshold; 287 } 288 289 } 290 291 /// Mark as used the sample record for the given function samples at 292 /// (LineOffset, Discriminator). 293 /// 294 /// \returns true if this is the first time we mark the given record. 295 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS, 296 uint32_t LineOffset, 297 uint32_t Discriminator, 298 uint64_t Samples) { 299 LineLocation Loc(LineOffset, Discriminator); 300 unsigned &Count = SampleCoverage[FS][Loc]; 301 bool FirstTime = (++Count == 1); 302 if (FirstTime) 303 TotalUsedSamples += Samples; 304 return FirstTime; 305 } 306 307 /// Return the number of sample records that were applied from this profile. 308 /// 309 /// This count does not include records from cold inlined callsites. 310 unsigned 311 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const { 312 auto I = SampleCoverage.find(FS); 313 314 // The size of the coverage map for FS represents the number of records 315 // that were marked used at least once. 316 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0; 317 318 // If there are inlined callsites in this function, count the samples found 319 // in the respective bodies. However, do not bother counting callees with 0 320 // total samples, these are callees that were never invoked at runtime. 321 for (const auto &I : FS->getCallsiteSamples()) { 322 const FunctionSamples *CalleeSamples = &I.second; 323 if (callsiteIsHot(FS, CalleeSamples)) 324 Count += countUsedRecords(CalleeSamples); 325 } 326 327 return Count; 328 } 329 330 /// Return the number of sample records in the body of this profile. 331 /// 332 /// This count does not include records from cold inlined callsites. 333 unsigned 334 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const { 335 unsigned Count = FS->getBodySamples().size(); 336 337 // Only count records in hot callsites. 338 for (const auto &I : FS->getCallsiteSamples()) { 339 const FunctionSamples *CalleeSamples = &I.second; 340 if (callsiteIsHot(FS, CalleeSamples)) 341 Count += countBodyRecords(CalleeSamples); 342 } 343 344 return Count; 345 } 346 347 /// Return the number of samples collected in the body of this profile. 348 /// 349 /// This count does not include samples from cold inlined callsites. 350 uint64_t 351 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const { 352 uint64_t Total = 0; 353 for (const auto &I : FS->getBodySamples()) 354 Total += I.second.getSamples(); 355 356 // Only count samples in hot callsites. 357 for (const auto &I : FS->getCallsiteSamples()) { 358 const FunctionSamples *CalleeSamples = &I.second; 359 if (callsiteIsHot(FS, CalleeSamples)) 360 Total += countBodySamples(CalleeSamples); 361 } 362 363 return Total; 364 } 365 366 /// Return the fraction of sample records used in this profile. 367 /// 368 /// The returned value is an unsigned integer in the range 0-100 indicating 369 /// the percentage of sample records that were used while applying this 370 /// profile to the associated function. 371 unsigned SampleCoverageTracker::computeCoverage(unsigned Used, 372 unsigned Total) const { 373 assert(Used <= Total && 374 "number of used records cannot exceed the total number of records"); 375 return Total > 0 ? Used * 100 / Total : 100; 376 } 377 378 /// Clear all the per-function data used to load samples and propagate weights. 379 void SampleProfileLoader::clearFunctionData() { 380 BlockWeights.clear(); 381 EdgeWeights.clear(); 382 VisitedBlocks.clear(); 383 VisitedEdges.clear(); 384 EquivalenceClass.clear(); 385 DT = nullptr; 386 PDT = nullptr; 387 LI = nullptr; 388 Predecessors.clear(); 389 Successors.clear(); 390 CoverageTracker.clear(); 391 } 392 393 /// \brief Returns the offset of lineno \p L to head_lineno \p H 394 /// 395 /// \param L Lineno 396 /// \param H Header lineno of the function 397 /// 398 /// \returns offset to the header lineno. 16 bits are used to represent offset. 399 /// We assume that a single function will not exceed 65535 LOC. 400 unsigned SampleProfileLoader::getOffset(unsigned L, unsigned H) const { 401 return (L - H) & 0xffff; 402 } 403 404 /// \brief Print the weight of edge \p E on stream \p OS. 405 /// 406 /// \param OS Stream to emit the output to. 407 /// \param E Edge to print. 408 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) { 409 OS << "weight[" << E.first->getName() << "->" << E.second->getName() 410 << "]: " << EdgeWeights[E] << "\n"; 411 } 412 413 /// \brief Print the equivalence class of block \p BB on stream \p OS. 414 /// 415 /// \param OS Stream to emit the output to. 416 /// \param BB Block to print. 417 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS, 418 const BasicBlock *BB) { 419 const BasicBlock *Equiv = EquivalenceClass[BB]; 420 OS << "equivalence[" << BB->getName() 421 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n"; 422 } 423 424 /// \brief Print the weight of block \p BB on stream \p OS. 425 /// 426 /// \param OS Stream to emit the output to. 427 /// \param BB Block to print. 428 void SampleProfileLoader::printBlockWeight(raw_ostream &OS, 429 const BasicBlock *BB) const { 430 const auto &I = BlockWeights.find(BB); 431 uint64_t W = (I == BlockWeights.end() ? 0 : I->second); 432 OS << "weight[" << BB->getName() << "]: " << W << "\n"; 433 } 434 435 /// \brief Get the weight for an instruction. 436 /// 437 /// The "weight" of an instruction \p Inst is the number of samples 438 /// collected on that instruction at runtime. To retrieve it, we 439 /// need to compute the line number of \p Inst relative to the start of its 440 /// function. We use HeaderLineno to compute the offset. We then 441 /// look up the samples collected for \p Inst using BodySamples. 442 /// 443 /// \param Inst Instruction to query. 444 /// 445 /// \returns the weight of \p Inst. 446 ErrorOr<uint64_t> 447 SampleProfileLoader::getInstWeight(const Instruction &Inst) const { 448 DebugLoc DLoc = Inst.getDebugLoc(); 449 if (!DLoc) 450 return std::error_code(); 451 452 const FunctionSamples *FS = findFunctionSamples(Inst); 453 if (!FS) 454 return std::error_code(); 455 456 const DILocation *DIL = DLoc; 457 unsigned Lineno = DLoc.getLine(); 458 unsigned HeaderLineno = DIL->getScope()->getSubprogram()->getLine(); 459 460 uint32_t LineOffset = getOffset(Lineno, HeaderLineno); 461 uint32_t Discriminator = DIL->getDiscriminator(); 462 ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator); 463 if (R) { 464 bool FirstMark = 465 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get()); 466 if (FirstMark) { 467 const Function *F = Inst.getParent()->getParent(); 468 LLVMContext &Ctx = F->getContext(); 469 emitOptimizationRemark( 470 Ctx, DEBUG_TYPE, *F, DLoc, 471 Twine("Applied ") + Twine(*R) + " samples from profile (offset: " + 472 Twine(LineOffset) + 473 ((Discriminator) ? Twine(".") + Twine(Discriminator) : "") + ")"); 474 } 475 DEBUG(dbgs() << " " << Lineno << "." << DIL->getDiscriminator() << ":" 476 << Inst << " (line offset: " << Lineno - HeaderLineno << "." 477 << DIL->getDiscriminator() << " - weight: " << R.get() 478 << ")\n"); 479 } 480 return R; 481 } 482 483 /// \brief Compute the weight of a basic block. 484 /// 485 /// The weight of basic block \p BB is the maximum weight of all the 486 /// instructions in BB. 487 /// 488 /// \param BB The basic block to query. 489 /// 490 /// \returns the weight for \p BB. 491 ErrorOr<uint64_t> 492 SampleProfileLoader::getBlockWeight(const BasicBlock *BB) const { 493 bool Found = false; 494 uint64_t Weight = 0; 495 for (auto &I : BB->getInstList()) { 496 const ErrorOr<uint64_t> &R = getInstWeight(I); 497 if (R && R.get() >= Weight) { 498 Weight = R.get(); 499 Found = true; 500 } 501 } 502 if (Found) 503 return Weight; 504 else 505 return std::error_code(); 506 } 507 508 /// \brief Compute and store the weights of every basic block. 509 /// 510 /// This populates the BlockWeights map by computing 511 /// the weights of every basic block in the CFG. 512 /// 513 /// \param F The function to query. 514 bool SampleProfileLoader::computeBlockWeights(Function &F) { 515 bool Changed = false; 516 DEBUG(dbgs() << "Block weights\n"); 517 for (const auto &BB : F) { 518 ErrorOr<uint64_t> Weight = getBlockWeight(&BB); 519 if (Weight) { 520 BlockWeights[&BB] = Weight.get(); 521 VisitedBlocks.insert(&BB); 522 Changed = true; 523 } 524 DEBUG(printBlockWeight(dbgs(), &BB)); 525 } 526 527 return Changed; 528 } 529 530 /// \brief Get the FunctionSamples for a call instruction. 531 /// 532 /// The FunctionSamples of a call instruction \p Inst is the inlined 533 /// instance in which that call instruction is calling to. It contains 534 /// all samples that resides in the inlined instance. We first find the 535 /// inlined instance in which the call instruction is from, then we 536 /// traverse its children to find the callsite with the matching 537 /// location and callee function name. 538 /// 539 /// \param Inst Call instruction to query. 540 /// 541 /// \returns The FunctionSamples pointer to the inlined instance. 542 const FunctionSamples * 543 SampleProfileLoader::findCalleeFunctionSamples(const CallInst &Inst) const { 544 const DILocation *DIL = Inst.getDebugLoc(); 545 if (!DIL) { 546 return nullptr; 547 } 548 DISubprogram *SP = DIL->getScope()->getSubprogram(); 549 if (!SP) 550 return nullptr; 551 552 Function *CalleeFunc = Inst.getCalledFunction(); 553 if (!CalleeFunc) { 554 return nullptr; 555 } 556 557 StringRef CalleeName = CalleeFunc->getName(); 558 const FunctionSamples *FS = findFunctionSamples(Inst); 559 if (FS == nullptr) 560 return nullptr; 561 562 return FS->findFunctionSamplesAt( 563 CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()), 564 DIL->getDiscriminator(), CalleeName)); 565 } 566 567 /// \brief Get the FunctionSamples for an instruction. 568 /// 569 /// The FunctionSamples of an instruction \p Inst is the inlined instance 570 /// in which that instruction is coming from. We traverse the inline stack 571 /// of that instruction, and match it with the tree nodes in the profile. 572 /// 573 /// \param Inst Instruction to query. 574 /// 575 /// \returns the FunctionSamples pointer to the inlined instance. 576 const FunctionSamples * 577 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const { 578 SmallVector<CallsiteLocation, 10> S; 579 const DILocation *DIL = Inst.getDebugLoc(); 580 if (!DIL) { 581 return Samples; 582 } 583 StringRef CalleeName; 584 for (const DILocation *DIL = Inst.getDebugLoc(); DIL; 585 DIL = DIL->getInlinedAt()) { 586 DISubprogram *SP = DIL->getScope()->getSubprogram(); 587 if (!SP) 588 return nullptr; 589 if (!CalleeName.empty()) { 590 S.push_back(CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()), 591 DIL->getDiscriminator(), CalleeName)); 592 } 593 CalleeName = SP->getLinkageName(); 594 } 595 if (S.size() == 0) 596 return Samples; 597 const FunctionSamples *FS = Samples; 598 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) { 599 FS = FS->findFunctionSamplesAt(S[i]); 600 } 601 return FS; 602 } 603 604 /// \brief Emit an inline hint if \p F is globally hot or cold. 605 /// 606 /// If \p F consumes a significant fraction of samples (indicated by 607 /// SampleProfileGlobalHotThreshold), apply the InlineHint attribute for the 608 /// inliner to consider the function hot. 609 /// 610 /// If \p F consumes a small fraction of samples (indicated by 611 /// SampleProfileGlobalColdThreshold), apply the Cold attribute for the inliner 612 /// to consider the function cold. 613 /// 614 /// FIXME - This setting of inline hints is sub-optimal. Instead of marking a 615 /// function globally hot or cold, we should be annotating individual callsites. 616 /// This is not currently possible, but work on the inliner will eventually 617 /// provide this ability. See http://reviews.llvm.org/D15003 for details and 618 /// discussion. 619 /// 620 /// \returns True if either attribute was applied to \p F. 621 bool SampleProfileLoader::emitInlineHints(Function &F) { 622 if (TotalCollectedSamples == 0) 623 return false; 624 625 uint64_t FunctionSamples = Samples->getTotalSamples(); 626 double SamplesPercent = 627 (double)FunctionSamples / (double)TotalCollectedSamples * 100.0; 628 629 // If the function collected more samples than the hot threshold, mark 630 // it globally hot. 631 if (SamplesPercent >= SampleProfileGlobalHotThreshold) { 632 F.addFnAttr(llvm::Attribute::InlineHint); 633 std::string Msg; 634 raw_string_ostream S(Msg); 635 S << "Applied inline hint to globally hot function '" << F.getName() 636 << "' with " << format("%.2f", SamplesPercent) 637 << "% of samples (threshold: " 638 << format("%.2f", SampleProfileGlobalHotThreshold.getValue()) << "%)"; 639 S.flush(); 640 emitOptimizationRemark(F.getContext(), DEBUG_TYPE, F, DebugLoc(), Msg); 641 return true; 642 } 643 644 // If the function collected fewer samples than the cold threshold, mark 645 // it globally cold. 646 if (SamplesPercent <= SampleProfileGlobalColdThreshold) { 647 F.addFnAttr(llvm::Attribute::Cold); 648 std::string Msg; 649 raw_string_ostream S(Msg); 650 S << "Applied cold hint to globally cold function '" << F.getName() 651 << "' with " << format("%.2f", SamplesPercent) 652 << "% of samples (threshold: " 653 << format("%.2f", SampleProfileGlobalColdThreshold.getValue()) << "%)"; 654 S.flush(); 655 emitOptimizationRemark(F.getContext(), DEBUG_TYPE, F, DebugLoc(), Msg); 656 return true; 657 } 658 659 return false; 660 } 661 662 /// \brief Iteratively inline hot callsites of a function. 663 /// 664 /// Iteratively traverse all callsites of the function \p F, and find if 665 /// the corresponding inlined instance exists and is hot in profile. If 666 /// it is hot enough, inline the callsites and adds new callsites of the 667 /// callee into the caller. 668 /// 669 /// TODO: investigate the possibility of not invoking InlineFunction directly. 670 /// 671 /// \param F function to perform iterative inlining. 672 /// 673 /// \returns True if there is any inline happened. 674 bool SampleProfileLoader::inlineHotFunctions(Function &F) { 675 bool Changed = false; 676 LLVMContext &Ctx = F.getContext(); 677 while (true) { 678 bool LocalChanged = false; 679 SmallVector<CallInst *, 10> CIS; 680 for (auto &BB : F) { 681 for (auto &I : BB.getInstList()) { 682 CallInst *CI = dyn_cast<CallInst>(&I); 683 if (CI && callsiteIsHot(Samples, findCalleeFunctionSamples(*CI))) 684 CIS.push_back(CI); 685 } 686 } 687 for (auto CI : CIS) { 688 InlineFunctionInfo IFI; 689 Function *CalledFunction = CI->getCalledFunction(); 690 DebugLoc DLoc = CI->getDebugLoc(); 691 uint64_t NumSamples = findCalleeFunctionSamples(*CI)->getTotalSamples(); 692 if (InlineFunction(CI, IFI)) { 693 LocalChanged = true; 694 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc, 695 Twine("inlined hot callee '") + 696 CalledFunction->getName() + "' with " + 697 Twine(NumSamples) + " samples into '" + 698 F.getName() + "'"); 699 } 700 } 701 if (LocalChanged) { 702 Changed = true; 703 } else { 704 break; 705 } 706 } 707 return Changed; 708 } 709 710 /// \brief Find equivalence classes for the given block. 711 /// 712 /// This finds all the blocks that are guaranteed to execute the same 713 /// number of times as \p BB1. To do this, it traverses all the 714 /// descendants of \p BB1 in the dominator or post-dominator tree. 715 /// 716 /// A block BB2 will be in the same equivalence class as \p BB1 if 717 /// the following holds: 718 /// 719 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2 720 /// is a descendant of \p BB1 in the dominator tree, then BB2 should 721 /// dominate BB1 in the post-dominator tree. 722 /// 723 /// 2- Both BB2 and \p BB1 must be in the same loop. 724 /// 725 /// For every block BB2 that meets those two requirements, we set BB2's 726 /// equivalence class to \p BB1. 727 /// 728 /// \param BB1 Block to check. 729 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree. 730 /// \param DomTree Opposite dominator tree. If \p Descendants is filled 731 /// with blocks from \p BB1's dominator tree, then 732 /// this is the post-dominator tree, and vice versa. 733 void SampleProfileLoader::findEquivalencesFor( 734 BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants, 735 DominatorTreeBase<BasicBlock> *DomTree) { 736 const BasicBlock *EC = EquivalenceClass[BB1]; 737 uint64_t Weight = BlockWeights[EC]; 738 for (const auto *BB2 : Descendants) { 739 bool IsDomParent = DomTree->dominates(BB2, BB1); 740 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2); 741 if (BB1 != BB2 && IsDomParent && IsInSameLoop) { 742 EquivalenceClass[BB2] = EC; 743 744 // If BB2 is heavier than BB1, make BB2 have the same weight 745 // as BB1. 746 // 747 // Note that we don't worry about the opposite situation here 748 // (when BB2 is lighter than BB1). We will deal with this 749 // during the propagation phase. Right now, we just want to 750 // make sure that BB1 has the largest weight of all the 751 // members of its equivalence set. 752 Weight = std::max(Weight, BlockWeights[BB2]); 753 } 754 } 755 BlockWeights[EC] = Weight; 756 } 757 758 /// \brief Find equivalence classes. 759 /// 760 /// Since samples may be missing from blocks, we can fill in the gaps by setting 761 /// the weights of all the blocks in the same equivalence class to the same 762 /// weight. To compute the concept of equivalence, we use dominance and loop 763 /// information. Two blocks B1 and B2 are in the same equivalence class if B1 764 /// dominates B2, B2 post-dominates B1 and both are in the same loop. 765 /// 766 /// \param F The function to query. 767 void SampleProfileLoader::findEquivalenceClasses(Function &F) { 768 SmallVector<BasicBlock *, 8> DominatedBBs; 769 DEBUG(dbgs() << "\nBlock equivalence classes\n"); 770 // Find equivalence sets based on dominance and post-dominance information. 771 for (auto &BB : F) { 772 BasicBlock *BB1 = &BB; 773 774 // Compute BB1's equivalence class once. 775 if (EquivalenceClass.count(BB1)) { 776 DEBUG(printBlockEquivalence(dbgs(), BB1)); 777 continue; 778 } 779 780 // By default, blocks are in their own equivalence class. 781 EquivalenceClass[BB1] = BB1; 782 783 // Traverse all the blocks dominated by BB1. We are looking for 784 // every basic block BB2 such that: 785 // 786 // 1- BB1 dominates BB2. 787 // 2- BB2 post-dominates BB1. 788 // 3- BB1 and BB2 are in the same loop nest. 789 // 790 // If all those conditions hold, it means that BB2 is executed 791 // as many times as BB1, so they are placed in the same equivalence 792 // class by making BB2's equivalence class be BB1. 793 DominatedBBs.clear(); 794 DT->getDescendants(BB1, DominatedBBs); 795 findEquivalencesFor(BB1, DominatedBBs, PDT.get()); 796 797 DEBUG(printBlockEquivalence(dbgs(), BB1)); 798 } 799 800 // Assign weights to equivalence classes. 801 // 802 // All the basic blocks in the same equivalence class will execute 803 // the same number of times. Since we know that the head block in 804 // each equivalence class has the largest weight, assign that weight 805 // to all the blocks in that equivalence class. 806 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n"); 807 for (auto &BI : F) { 808 const BasicBlock *BB = &BI; 809 const BasicBlock *EquivBB = EquivalenceClass[BB]; 810 if (BB != EquivBB) 811 BlockWeights[BB] = BlockWeights[EquivBB]; 812 DEBUG(printBlockWeight(dbgs(), BB)); 813 } 814 } 815 816 /// \brief Visit the given edge to decide if it has a valid weight. 817 /// 818 /// If \p E has not been visited before, we copy to \p UnknownEdge 819 /// and increment the count of unknown edges. 820 /// 821 /// \param E Edge to visit. 822 /// \param NumUnknownEdges Current number of unknown edges. 823 /// \param UnknownEdge Set if E has not been visited before. 824 /// 825 /// \returns E's weight, if known. Otherwise, return 0. 826 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges, 827 Edge *UnknownEdge) { 828 if (!VisitedEdges.count(E)) { 829 (*NumUnknownEdges)++; 830 *UnknownEdge = E; 831 return 0; 832 } 833 834 return EdgeWeights[E]; 835 } 836 837 /// \brief Propagate weights through incoming/outgoing edges. 838 /// 839 /// If the weight of a basic block is known, and there is only one edge 840 /// with an unknown weight, we can calculate the weight of that edge. 841 /// 842 /// Similarly, if all the edges have a known count, we can calculate the 843 /// count of the basic block, if needed. 844 /// 845 /// \param F Function to process. 846 /// 847 /// \returns True if new weights were assigned to edges or blocks. 848 bool SampleProfileLoader::propagateThroughEdges(Function &F) { 849 bool Changed = false; 850 DEBUG(dbgs() << "\nPropagation through edges\n"); 851 for (const auto &BI : F) { 852 const BasicBlock *BB = &BI; 853 const BasicBlock *EC = EquivalenceClass[BB]; 854 855 // Visit all the predecessor and successor edges to determine 856 // which ones have a weight assigned already. Note that it doesn't 857 // matter that we only keep track of a single unknown edge. The 858 // only case we are interested in handling is when only a single 859 // edge is unknown (see setEdgeOrBlockWeight). 860 for (unsigned i = 0; i < 2; i++) { 861 uint64_t TotalWeight = 0; 862 unsigned NumUnknownEdges = 0; 863 Edge UnknownEdge, SelfReferentialEdge; 864 865 if (i == 0) { 866 // First, visit all predecessor edges. 867 for (auto *Pred : Predecessors[BB]) { 868 Edge E = std::make_pair(Pred, BB); 869 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge); 870 if (E.first == E.second) 871 SelfReferentialEdge = E; 872 } 873 } else { 874 // On the second round, visit all successor edges. 875 for (auto *Succ : Successors[BB]) { 876 Edge E = std::make_pair(BB, Succ); 877 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge); 878 } 879 } 880 881 // After visiting all the edges, there are three cases that we 882 // can handle immediately: 883 // 884 // - All the edge weights are known (i.e., NumUnknownEdges == 0). 885 // In this case, we simply check that the sum of all the edges 886 // is the same as BB's weight. If not, we change BB's weight 887 // to match. Additionally, if BB had not been visited before, 888 // we mark it visited. 889 // 890 // - Only one edge is unknown and BB has already been visited. 891 // In this case, we can compute the weight of the edge by 892 // subtracting the total block weight from all the known 893 // edge weights. If the edges weight more than BB, then the 894 // edge of the last remaining edge is set to zero. 895 // 896 // - There exists a self-referential edge and the weight of BB is 897 // known. In this case, this edge can be based on BB's weight. 898 // We add up all the other known edges and set the weight on 899 // the self-referential edge as we did in the previous case. 900 // 901 // In any other case, we must continue iterating. Eventually, 902 // all edges will get a weight, or iteration will stop when 903 // it reaches SampleProfileMaxPropagateIterations. 904 if (NumUnknownEdges <= 1) { 905 uint64_t &BBWeight = BlockWeights[EC]; 906 if (NumUnknownEdges == 0) { 907 // If we already know the weight of all edges, the weight of the 908 // basic block can be computed. It should be no larger than the sum 909 // of all edge weights. 910 if (TotalWeight > BBWeight) { 911 BBWeight = TotalWeight; 912 Changed = true; 913 DEBUG(dbgs() << "All edge weights for " << BB->getName() 914 << " known. Set weight for block: "; 915 printBlockWeight(dbgs(), BB);); 916 } 917 if (VisitedBlocks.insert(EC).second) 918 Changed = true; 919 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) { 920 // If there is a single unknown edge and the block has been 921 // visited, then we can compute E's weight. 922 if (BBWeight >= TotalWeight) 923 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight; 924 else 925 EdgeWeights[UnknownEdge] = 0; 926 VisitedEdges.insert(UnknownEdge); 927 Changed = true; 928 DEBUG(dbgs() << "Set weight for edge: "; 929 printEdgeWeight(dbgs(), UnknownEdge)); 930 } 931 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) { 932 uint64_t &BBWeight = BlockWeights[BB]; 933 // We have a self-referential edge and the weight of BB is known. 934 if (BBWeight >= TotalWeight) 935 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight; 936 else 937 EdgeWeights[SelfReferentialEdge] = 0; 938 VisitedEdges.insert(SelfReferentialEdge); 939 Changed = true; 940 DEBUG(dbgs() << "Set self-referential edge weight to: "; 941 printEdgeWeight(dbgs(), SelfReferentialEdge)); 942 } 943 } 944 } 945 946 return Changed; 947 } 948 949 /// \brief Build in/out edge lists for each basic block in the CFG. 950 /// 951 /// We are interested in unique edges. If a block B1 has multiple 952 /// edges to another block B2, we only add a single B1->B2 edge. 953 void SampleProfileLoader::buildEdges(Function &F) { 954 for (auto &BI : F) { 955 BasicBlock *B1 = &BI; 956 957 // Add predecessors for B1. 958 SmallPtrSet<BasicBlock *, 16> Visited; 959 if (!Predecessors[B1].empty()) 960 llvm_unreachable("Found a stale predecessors list in a basic block."); 961 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) { 962 BasicBlock *B2 = *PI; 963 if (Visited.insert(B2).second) 964 Predecessors[B1].push_back(B2); 965 } 966 967 // Add successors for B1. 968 Visited.clear(); 969 if (!Successors[B1].empty()) 970 llvm_unreachable("Found a stale successors list in a basic block."); 971 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) { 972 BasicBlock *B2 = *SI; 973 if (Visited.insert(B2).second) 974 Successors[B1].push_back(B2); 975 } 976 } 977 } 978 979 /// \brief Propagate weights into edges 980 /// 981 /// The following rules are applied to every block BB in the CFG: 982 /// 983 /// - If BB has a single predecessor/successor, then the weight 984 /// of that edge is the weight of the block. 985 /// 986 /// - If all incoming or outgoing edges are known except one, and the 987 /// weight of the block is already known, the weight of the unknown 988 /// edge will be the weight of the block minus the sum of all the known 989 /// edges. If the sum of all the known edges is larger than BB's weight, 990 /// we set the unknown edge weight to zero. 991 /// 992 /// - If there is a self-referential edge, and the weight of the block is 993 /// known, the weight for that edge is set to the weight of the block 994 /// minus the weight of the other incoming edges to that block (if 995 /// known). 996 void SampleProfileLoader::propagateWeights(Function &F) { 997 bool Changed = true; 998 unsigned I = 0; 999 1000 // Add an entry count to the function using the samples gathered 1001 // at the function entry. 1002 F.setEntryCount(Samples->getHeadSamples()); 1003 1004 // Before propagation starts, build, for each block, a list of 1005 // unique predecessors and successors. This is necessary to handle 1006 // identical edges in multiway branches. Since we visit all blocks and all 1007 // edges of the CFG, it is cleaner to build these lists once at the start 1008 // of the pass. 1009 buildEdges(F); 1010 1011 // Propagate until we converge or we go past the iteration limit. 1012 while (Changed && I++ < SampleProfileMaxPropagateIterations) { 1013 Changed = propagateThroughEdges(F); 1014 } 1015 1016 // Generate MD_prof metadata for every branch instruction using the 1017 // edge weights computed during propagation. 1018 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n"); 1019 LLVMContext &Ctx = F.getContext(); 1020 MDBuilder MDB(Ctx); 1021 for (auto &BI : F) { 1022 BasicBlock *BB = &BI; 1023 TerminatorInst *TI = BB->getTerminator(); 1024 if (TI->getNumSuccessors() == 1) 1025 continue; 1026 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI)) 1027 continue; 1028 1029 DEBUG(dbgs() << "\nGetting weights for branch at line " 1030 << TI->getDebugLoc().getLine() << ".\n"); 1031 SmallVector<uint32_t, 4> Weights; 1032 uint32_t MaxWeight = 0; 1033 DebugLoc MaxDestLoc; 1034 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) { 1035 BasicBlock *Succ = TI->getSuccessor(I); 1036 Edge E = std::make_pair(BB, Succ); 1037 uint64_t Weight = EdgeWeights[E]; 1038 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E)); 1039 // Use uint32_t saturated arithmetic to adjust the incoming weights, 1040 // if needed. Sample counts in profiles are 64-bit unsigned values, 1041 // but internally branch weights are expressed as 32-bit values. 1042 if (Weight > std::numeric_limits<uint32_t>::max()) { 1043 DEBUG(dbgs() << " (saturated due to uint32_t overflow)"); 1044 Weight = std::numeric_limits<uint32_t>::max(); 1045 } 1046 Weights.push_back(static_cast<uint32_t>(Weight)); 1047 if (Weight != 0) { 1048 if (Weight > MaxWeight) { 1049 MaxWeight = Weight; 1050 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc(); 1051 } 1052 } 1053 } 1054 1055 // Only set weights if there is at least one non-zero weight. 1056 // In any other case, let the analyzer set weights. 1057 if (MaxWeight > 0) { 1058 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n"); 1059 TI->setMetadata(llvm::LLVMContext::MD_prof, 1060 MDB.createBranchWeights(Weights)); 1061 DebugLoc BranchLoc = TI->getDebugLoc(); 1062 emitOptimizationRemark( 1063 Ctx, DEBUG_TYPE, F, MaxDestLoc, 1064 Twine("most popular destination for conditional branches at ") + 1065 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" + 1066 Twine(BranchLoc.getLine()) + ":" + 1067 Twine(BranchLoc.getCol())) 1068 : Twine("<UNKNOWN LOCATION>"))); 1069 } else { 1070 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n"); 1071 } 1072 } 1073 } 1074 1075 /// \brief Get the line number for the function header. 1076 /// 1077 /// This looks up function \p F in the current compilation unit and 1078 /// retrieves the line number where the function is defined. This is 1079 /// line 0 for all the samples read from the profile file. Every line 1080 /// number is relative to this line. 1081 /// 1082 /// \param F Function object to query. 1083 /// 1084 /// \returns the line number where \p F is defined. If it returns 0, 1085 /// it means that there is no debug information available for \p F. 1086 unsigned SampleProfileLoader::getFunctionLoc(Function &F) { 1087 if (DISubprogram *S = getDISubprogram(&F)) 1088 return S->getLine(); 1089 1090 // If the start of \p F is missing, emit a diagnostic to inform the user 1091 // about the missed opportunity. 1092 F.getContext().diagnose(DiagnosticInfoSampleProfile( 1093 "No debug information found in function " + F.getName() + 1094 ": Function profile not used", 1095 DS_Warning)); 1096 return 0; 1097 } 1098 1099 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) { 1100 DT.reset(new DominatorTree); 1101 DT->recalculate(F); 1102 1103 PDT.reset(new DominatorTreeBase<BasicBlock>(true)); 1104 PDT->recalculate(F); 1105 1106 LI.reset(new LoopInfo); 1107 LI->analyze(*DT); 1108 } 1109 1110 /// \brief Generate branch weight metadata for all branches in \p F. 1111 /// 1112 /// Branch weights are computed out of instruction samples using a 1113 /// propagation heuristic. Propagation proceeds in 3 phases: 1114 /// 1115 /// 1- Assignment of block weights. All the basic blocks in the function 1116 /// are initial assigned the same weight as their most frequently 1117 /// executed instruction. 1118 /// 1119 /// 2- Creation of equivalence classes. Since samples may be missing from 1120 /// blocks, we can fill in the gaps by setting the weights of all the 1121 /// blocks in the same equivalence class to the same weight. To compute 1122 /// the concept of equivalence, we use dominance and loop information. 1123 /// Two blocks B1 and B2 are in the same equivalence class if B1 1124 /// dominates B2, B2 post-dominates B1 and both are in the same loop. 1125 /// 1126 /// 3- Propagation of block weights into edges. This uses a simple 1127 /// propagation heuristic. The following rules are applied to every 1128 /// block BB in the CFG: 1129 /// 1130 /// - If BB has a single predecessor/successor, then the weight 1131 /// of that edge is the weight of the block. 1132 /// 1133 /// - If all the edges are known except one, and the weight of the 1134 /// block is already known, the weight of the unknown edge will 1135 /// be the weight of the block minus the sum of all the known 1136 /// edges. If the sum of all the known edges is larger than BB's weight, 1137 /// we set the unknown edge weight to zero. 1138 /// 1139 /// - If there is a self-referential edge, and the weight of the block is 1140 /// known, the weight for that edge is set to the weight of the block 1141 /// minus the weight of the other incoming edges to that block (if 1142 /// known). 1143 /// 1144 /// Since this propagation is not guaranteed to finalize for every CFG, we 1145 /// only allow it to proceed for a limited number of iterations (controlled 1146 /// by -sample-profile-max-propagate-iterations). 1147 /// 1148 /// FIXME: Try to replace this propagation heuristic with a scheme 1149 /// that is guaranteed to finalize. A work-list approach similar to 1150 /// the standard value propagation algorithm used by SSA-CCP might 1151 /// work here. 1152 /// 1153 /// Once all the branch weights are computed, we emit the MD_prof 1154 /// metadata on BB using the computed values for each of its branches. 1155 /// 1156 /// \param F The function to query. 1157 /// 1158 /// \returns true if \p F was modified. Returns false, otherwise. 1159 bool SampleProfileLoader::emitAnnotations(Function &F) { 1160 bool Changed = false; 1161 1162 if (getFunctionLoc(F) == 0) 1163 return false; 1164 1165 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName() 1166 << ": " << getFunctionLoc(F) << "\n"); 1167 1168 Changed |= emitInlineHints(F); 1169 1170 Changed |= inlineHotFunctions(F); 1171 1172 // Compute basic block weights. 1173 Changed |= computeBlockWeights(F); 1174 1175 if (Changed) { 1176 // Compute dominance and loop info needed for propagation. 1177 computeDominanceAndLoopInfo(F); 1178 1179 // Find equivalence classes. 1180 findEquivalenceClasses(F); 1181 1182 // Propagate weights to all edges. 1183 propagateWeights(F); 1184 } 1185 1186 // If coverage checking was requested, compute it now. 1187 if (SampleProfileRecordCoverage) { 1188 unsigned Used = CoverageTracker.countUsedRecords(Samples); 1189 unsigned Total = CoverageTracker.countBodyRecords(Samples); 1190 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total); 1191 if (Coverage < SampleProfileRecordCoverage) { 1192 F.getContext().diagnose(DiagnosticInfoSampleProfile( 1193 getDISubprogram(&F)->getFilename(), getFunctionLoc(F), 1194 Twine(Used) + " of " + Twine(Total) + " available profile records (" + 1195 Twine(Coverage) + "%) were applied", 1196 DS_Warning)); 1197 } 1198 } 1199 1200 if (SampleProfileSampleCoverage) { 1201 uint64_t Used = CoverageTracker.getTotalUsedSamples(); 1202 uint64_t Total = CoverageTracker.countBodySamples(Samples); 1203 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total); 1204 if (Coverage < SampleProfileSampleCoverage) { 1205 F.getContext().diagnose(DiagnosticInfoSampleProfile( 1206 getDISubprogram(&F)->getFilename(), getFunctionLoc(F), 1207 Twine(Used) + " of " + Twine(Total) + " available profile samples (" + 1208 Twine(Coverage) + "%) were applied", 1209 DS_Warning)); 1210 } 1211 } 1212 return Changed; 1213 } 1214 1215 char SampleProfileLoader::ID = 0; 1216 INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile", 1217 "Sample Profile loader", false, false) 1218 INITIALIZE_PASS_DEPENDENCY(AddDiscriminators) 1219 INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile", 1220 "Sample Profile loader", false, false) 1221 1222 bool SampleProfileLoader::doInitialization(Module &M) { 1223 auto &Ctx = M.getContext(); 1224 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx); 1225 if (std::error_code EC = ReaderOrErr.getError()) { 1226 std::string Msg = "Could not open profile: " + EC.message(); 1227 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg)); 1228 return false; 1229 } 1230 Reader = std::move(ReaderOrErr.get()); 1231 ProfileIsValid = (Reader->read() == sampleprof_error::success); 1232 return true; 1233 } 1234 1235 ModulePass *llvm::createSampleProfileLoaderPass() { 1236 return new SampleProfileLoader(SampleProfileFile); 1237 } 1238 1239 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) { 1240 return new SampleProfileLoader(Name); 1241 } 1242 1243 bool SampleProfileLoader::runOnModule(Module &M) { 1244 if (!ProfileIsValid) 1245 return false; 1246 1247 // Compute the total number of samples collected in this profile. 1248 for (const auto &I : Reader->getProfiles()) 1249 TotalCollectedSamples += I.second.getTotalSamples(); 1250 1251 bool retval = false; 1252 for (auto &F : M) 1253 if (!F.isDeclaration()) { 1254 clearFunctionData(); 1255 retval |= runOnFunction(F); 1256 } 1257 return retval; 1258 } 1259 1260 bool SampleProfileLoader::runOnFunction(Function &F) { 1261 Samples = Reader->getSamplesFor(F); 1262 if (!Samples->empty()) 1263 return emitAnnotations(F); 1264 return false; 1265 } 1266